--core-offset has no effect

[soria-aurelien]

Hello,
I tried to underclock my RTX 5090 (With Nvidia Open Kernel modules) using your scripts but it does not seem to work as intended.

Here is the default curve of the card on MSI Afterburner

The underclock I'm trying to achieve is the following one (Working on windows with MSI Afterburner) :

An offset of 1000 to reach 2317Mhz @ 960mv with of transition clock of 900Mhz.
When applying
sudo python3 nvml-undervolt.py --core-offset 1000 --target-clock 2317 --transition-clock 900
When starting Portal RTX, the clock of my RTX 5090 is locked at 900 (or any other value of transition-clock i set) .Modifying the --core-offset does not seem to change anything.

Unfortunately I can't check the voltage via nvidia-smi to check if there is an effective undervolt:

Attached GPUs                             : 1
GPU 00000000:01:00.0
    Voltage
        Graphics                          : N/A

[jacklul]

Might be a case of newer card not being completely supported by the NVML API, are you able to change the clock using nvidia-smi? Or the way PStates work with newer cards might be different

I'm no longer working on this script because sadly I'm back to Windows

[soria-aurelien]

That's a good guess !

Default clock:

Clocks
    Graphics                          : 30 MHz
    SM                                : 30 MHz
    Memory                            : 405 MHz
    Video                             : 637 MHz

The clock after running sudo nvidia-smi --lock-gpu-clocks=2317

Clocks
    Graphics                          : 288 MHz
    SM                                : 288 MHz
    Memory                            : 810 MHz
    Video                             : 1995 MHz

It might indeed be the source of the issue. Thank for your help 💯

No worry, I will look into it on my side and if I'm not able to fix it, I will use a power limit instead.

[jacklul]

The clock after running sudo nvidia-smi --lock-gpu-clocks=2317

Try sudo nvidia-smi --lock-gpu-clocks=2317,2317 instead
Might also need that --mode=1 parameter

According to this it is how it should be used mainly, while the singular clock is still allowed

But yeah, if nvidia-smi doesn't work then don't expect NVML API to work either, it's NVIDIA after all

[soria-aurelien]

Unfortunately, same result with sudo nvidia-smi --lock-gpu-clocks=2317,2317 --mode=1
I think I will need to wait a bit. The 5xxx series is still quite fresh and the drivers might be better over time.